Skip to content

When Will AI Pay for Itself?

What AI costs, and what it is starting to give back.

Sources and method

How every figure on this page was found, checked and labelled, and the full list of where each one comes from.

Evidence 1.1 · last checked 23 September 2026 · 304 references

How the evidence was gathered

Four research passes - two on costs (today's costs; cost curves and the railway history) and two on benefits - each with the same evidence standard: figure, unit, year, source, date, URL and a verification tag, with projections and vendor claims labelled. The cost dossiers tag each finding as confirmed in two independent sources, resting on one primary source, or unverified; unverified findings stay off the page. The benefit dossiers class each case as peer-reviewed or independently evaluated, a real deployment with reported numbers, or an estimate or projection. A fifth pass, on the social goals, covered both columns - poverty, schooling, work, discrimination and abuse imagery - to the same standard.

Cite this page

Houghton, J. (2026). When Will AI Pay for Itself? What AI costs, and what it is starting to give back (Version 1.1). https://when-will-ai-pay.vercel.app

Reading the badges

Two sources: Two independent organisations agree.
Two independent organisations agree.
One source: A single primary document, usually an official statistic or a peer-reviewed paper.
A single primary document, usually an official statistic or a peer-reviewed paper.
Vendor: The company's own figure, unaudited.
The company's own figure, unaudited.
Projection: A scenario, not a measurement.
A scenario, not a measurement.

Benefit rows also carry an evidence tier

(a) peer-reviewed
Randomised, peer-reviewed or independently evaluated.
(b) real deployment
A real deployment with organisation-reported numbers.
(c) estimate
An estimate or projection.

What changed in the evidence

  1. 1.1

    23 September 2026

    Six new ledger cards on the social goals - social protection, learning and work on the benefits side; jobs, discrimination and abuse imagery on the costs side - from a fifth research dossier, checked against source on 23 September 2026. Earlier figures were not re-checked in this release.

  2. 1.0

    22 September 2026

    First full set of figures, synthesised from the four research dossiers and checked against source.

Changes to the site itself, rather than its figures, are on the Progress tab.

Spotted an error, or know a better source? The Feedback button at the foot of every page comes straight to Joe.

References

304 works from 5 research dossiers · APA 7, alphabetical

  • Yao, X., Rushlow, D. R., Inselman, J. W., McCoy, R. G., Thacher, T. D., Behnken, E. M., ... Noseworthy, P. A. (2021). Artificial intelligence-enabled electrocardiograms for identification of patients with low ejection fraction: A pragmatic, randomized clinical trial. Nature Medicine, 27(5), 815-819. https://doi.org/10.1038/s41591-021-01335-4 (opens in a new tab)